Process-oriented machine learning for robust crop modeling under climate extremes

Agricultural systems are increasingly threatened by climate extremes such as heat waves, droughts, and compound events. Current crop modelling approaches face critical limitations: mechanistic models are physically consistent and transferable but computationally expensive and often fail to capture yield anomalies under extremes, while machine learning (ML) models achieve high predictive accuracy in-sample but lack robustness and explainability under out-of-distribution conditions. The proposed project, HybridCrop, addresses these weaknesses by pioneering a new class of hybrid, process-oriented crop models that integrate physics-informed ML modules with established mechanistic frameworks (EPIC, LPJmL). HybridCrop pursues four main objectives: 1) Development of physics-informed ML modules for selected processes (leaf area index, phenology, evapotranspiration, soil moisture, and yield) trained on a harmonized database comprising model simulations, Earth observation (EO) and in-situ data 2) Systematic integration of the new ML modules within LPJmL and EPIC as well as in standalone couplers. Evaluation will follow a parallel, stepwise workflow, contrasting hybrid, mechanistic, and observation-driven simulations to derive robust conclusions on the utility and limitations of ML modules. 3) Assessing the computational costs of the new hybrid models 4) Improving the scientific understanding of extreme event impacts on crop production, making use of high-resolution Single Model Initial-condition Large Ensembles (SMILEs). The project implementation is organized in ten work packages, ranging from data collection (WP1), module development (WPs 2–5), model coupling (WPs 6–7), benchmarking (WP8), application with SMILEs (WP9), to dissemination and open-source release (WP10). The project brings together internationally recognized experts in crop modelling, climate impacts, and machine learning, with proven track records in both mechanistic and ML approaches as well as in the development of open datasets and community standards. HybridCrop will exceed the current state of the art in several ways: The new combination of process-oriented ML modules with mechanistic approaches shows great potential for the reproduction of yield anomalies, larger model stability compared to existing ML approaches, and smaller computational costs compared to pure mechanistic models. The project will openly release benchmark datasets and hybrid model code, thus establishing a new methodological frontier for the AgMIP, GGCMI, and AgML communities. Beyond advancing scientific knowledge, its outcomes will directly inform adaptation planning, food security assessments, and international assessments such as the IPCC AR7. Ultimately, the project will position European research at the forefront of developing robust and efficient crop models that can capture the challenges of a more extreme climate future.

Options